[Feature]: support Multi-Scale-DeformAttention in deformable-detr (#878)

* add c++ ms_deform_atten

* fix cpp lint

* fix cpp lint

* clang format

* remove cmakefile

* google style

* clang-format precommit

* use clang-format-lint-action

* add transformer base class

* add merge

* add docstr

* add pyargs

* fix according to commments

* resiger module

* change to use basemodule

* add _ between build function

* split the name

* fix according to comments

* fix lint and fix unitest

* fix cpp lint

* fix bug of deformdetr_atten

* fix drop out

* fix residual

* use CUDA_1D_KERNEL_LOOP

mmcv/cnn/bricks/transformer.py

@@ -1,11 +1,15 @@
 import copy
+import math
 import warnings
 
 import torch
 import torch.nn as nn
 
 from mmcv import ConfigDict
-from mmcv.cnn import Linear, build_activation_layer, build_norm_layer
+from mmcv.cnn import (Linear, build_activation_layer, build_norm_layer,
+                      constant_init, xavier_init)
+from mmcv.ops.multi_scale_deform_attn import (
+    MultiScaleDeformableAttnFunction, multi_scale_deformable_attn_pytorch)
 from mmcv.runner.base_module import BaseModule
 from mmcv.utils import build_from_cfg
 from .registry import (ATTENTION, POSITIONAL_ENCODING, TRANSFORMER_LAYER,
@@ -135,6 +139,201 @@ def forward(self,
         return residual + self.dropout(out)
 
 
+@ATTENTION.register_module()
+class MultiScaleDeformableAttention(BaseModule):
+    """An attention module used in Deformable-Detr. `Deformable DETR:
+    Deformable Transformers for End-to-End Object Detection.
+
+      <https://arxiv.org/pdf/2010.04159.pdf>`_.
+
+    Args:
+        embed_dims (int): The embedding dimension of Attention.
+            Default: 256.
+        num_heads (int): Parallel attention heads. Default: 64.
+        num_levels (int): The number of feature map used in
+            Attention. Default: 4.
+        num_points (int): The number of sampling points for
+            each query in each head. Default: 4.
+        im2col_step (int): The step used in image_to_column.
+            Default: 64.
+        dropout (float): A Dropout layer on `inp_residual`.
+            Default: 0..
+        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
+            Default: None.
+    """
+
+    def __init__(self,
+                 embed_dims=256,
+                 num_heads=8,
+                 num_levels=4,
+                 num_points=4,
+                 im2col_step=64,
+                 dropout=0.1,
+                 norm_cfg=None,
+                 init_cfg=None):
+        super().__init__(init_cfg)
+        if embed_dims % num_heads != 0:
+            raise ValueError(f'embed_dims must be divisible by num_heads, '
+                             f'but got {embed_dims} and {num_heads}')
+        dim_per_head = embed_dims // num_heads
+        self.norm_cfg = norm_cfg
+        self.init_cfg = init_cfg
+        self.dropout = nn.Dropout(dropout)
+
+        # you'd better set dim_per_head to a power of 2
+        # which is more efficient in the CUDA implementation
+        def _is_power_of_2(n):
+            if (not isinstance(n, int)) or (n < 0):
+                raise ValueError(
+                    'invalid input for _is_power_of_2: {} (type: {})'.format(
+                        n, type(n)))
+            return (n & (n - 1) == 0) and n != 0
+
+        if not _is_power_of_2(dim_per_head):
+            warnings.warn(
+                "You'd better set embed_dims in "
+                'MultiScaleDeformAttention to make '
+                'the dimension of each attention head a power of 2 '
+                'which is more efficient in our CUDA implementation.')
+
+        self.im2col_step = im2col_step
+        self.embed_dims = embed_dims
+        self.num_levels = num_levels
+        self.num_heads = num_heads
+        self.num_points = num_points
+        self.sampling_offsets = nn.Linear(
+            embed_dims, num_heads * num_levels * num_points * 2)
+        self.attention_weights = nn.Linear(embed_dims,
+                                           num_heads * num_levels * num_points)
+        self.value_proj = nn.Linear(embed_dims, embed_dims)
+        self.output_proj = nn.Linear(embed_dims, embed_dims)
+        self.init_weight()
+
+    def init_weight(self):
+        """Default initialization for Parameters of Module."""
+        constant_init(self.sampling_offsets, 0.)
+        thetas = torch.arange(
+            self.num_heads,
+            dtype=torch.float32) * (2.0 * math.pi / self.num_heads)
+        grid_init = torch.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = (grid_init /
+                     grid_init.abs().max(-1, keepdim=True)[0]).view(
+                         self.num_heads, 1, 1,
+                         2).repeat(1, self.num_levels, self.num_points, 1)
+        for i in range(self.num_points):
+            grid_init[:, :, i, :] *= i + 1
+
+        self.sampling_offsets.bias.data = grid_init.view(-1)
+        constant_init(self.attention_weights, val=0., bias=0.)
+        xavier_init(self.value_proj, distribution='uniform', bias=0.)
+        xavier_init(self.output_proj, distribution='uniform', bias=0.)
+
+    def forward(self,
+                query,
+                key,
+                value,
+                residual=None,
+                query_pos=None,
+                key_padding_mask=None,
+                reference_points=None,
+                spatial_shapes=None,
+                level_start_index=None,
+                **kwargs):
+        """Forward Function of MultiScaleDeformAttention.
+
+        Args:
+            query (Tensor): Query of Transformer with shape
+                (num_query, bs, embed_dims).
+            key (Tensor): The key tensor with shape
+                `(num_key, bs, embed_dims)`.
+            value (Tensor): The value tensor with shape
+                `(num_key, bs, embed_dims)`.
+            residual (Tensor): The tensor used for addition, with the
+                same shape as `x`. Default None. If None, `x` will be used.
+            query_pos (Tensor): The positional encoding for `query`.
+                Default: None.
+            key_pos (Tensor): The positional encoding for `key`. Default
+                None.
+            reference_points (Tensor):  The normalized reference
+                points with shape (bs, num_query, num_levels, 2),
+                all elements is range in [0, 1], top-left (0,0),
+                bottom-right (1, 1), including padding area.
+                or (N, Length_{query}, num_levels, 4), add
+                additional two dimensions is (w, h) to
+                form reference boxes.
+            key_padding_mask (Tensor): ByteTensor for `query`, with
+                shape [bs, num_key].
+            spatial_shapes (Tensor): Spatial shape of features in
+                different level. With shape  (num_levels, 2),
+                last dimension represent (h, w).
+            level_start_index (Tensor): The start index of each level.
+                A tensor has shape (num_levels) and can be represented
+                as [0, h_0*w_0, h_0*w_0+h_1*w_1, ...].
+
+        Returns:
+             Tensor: forwarded results with shape [num_query, bs, embed_dims].
+        """
+
+        if key is None:
+            key = query
+        if value is None:
+            value = key
+
+        if residual is None:
+            inp_residual = query
+        if query_pos is not None:
+            query = query + query_pos
+
+        # change to (bs, num_query ,embed_dims)
+        query = query.permute(1, 0, 2)
+        value = value.permute(1, 0, 2)
+
+        bs, num_query, _ = query.shape
+        bs, num_key, _ = value.shape
+        assert (spatial_shapes[:, 0] * spatial_shapes[:, 1]).sum() == num_key
+
+        value = self.value_proj(value)
+        if key_padding_mask is not None:
+            value = value.masked_fill(key_padding_mask[..., None], 0.0)
+        value = value.view(bs, num_key, self.num_heads, -1)
+        sampling_offsets = self.sampling_offsets(query).view(
+            bs, num_query, self.num_heads, self.num_levels, self.num_points, 2)
+        attention_weights = self.attention_weights(query).view(
+            bs, num_query, self.num_heads, self.num_levels * self.num_points)
+        attention_weights = attention_weights.softmax(-1)
+
+        attention_weights = attention_weights.view(bs, num_query,
+                                                   self.num_heads,
+                                                   self.num_levels,
+                                                   self.num_points)
+        if reference_points.shape[-1] == 2:
+            offset_normalizer = torch.stack(
+                [spatial_shapes[..., 1], spatial_shapes[..., 0]], -1)
+            sampling_locations = reference_points[:, :, None, :, None, :] \
+                + sampling_offsets \
+                / offset_normalizer[None, None, None, :, None, :]
+        elif reference_points.shape[-1] == 4:
+            sampling_locations = reference_points[:, :, None, :, None, :2] \
+                + sampling_offsets / self.num_points \
+                * reference_points[:, :, None, :, None, 2:] \
+                * 0.5
+        else:
+            raise ValueError(
+                f'Last dim of reference_points must be'
+                f' 2 or 4, but get {reference_points.shape[-1]} instead.')
+        if torch.cuda.is_available():
+            output = MultiScaleDeformableAttnFunction.apply(
+                value, spatial_shapes, level_start_index, sampling_locations,
+                attention_weights, self.im2col_step)
+        else:
+            output = multi_scale_deformable_attn_pytorch(
+                value, spatial_shapes, level_start_index, sampling_locations,
+                attention_weights, self.im2col_step)
+        output = self.output_proj(output).permute(1, 0, 2)
+        # (num_query, bs ,embed_dims)
+        return self.dropout(output) + inp_residual
+
+
 class FFN(BaseModule):
     """Implements feed-forward networks (FFNs) with residual connection.